Automated Plant Disease Detection via ConvNeXt-Based Deep Learning and Progressive Transfer Learning

Abstract

Plant diseases cause the losses of up to 40% of global harvests and more than 220 billion USD to the world economy annually. These losses primarily affect small-scale farmers, who typically lack the resources to obtain expert-level plant disease diagnosis, highlighting the need for timely and reliable detection mechanisms. We propose an automated plant disease detection system based on the computer vision paradigm and ConvNeXt Base architecture that incorporates state-of-the-art hybrid transformer designs into the convolutional neural network through progressive fine-tuning. Our model achieves 94.22% top-1 and 97.83% top-3 accuracy on the custom dataset of 8,250 images of 15 classes (5 crops with 3 classes each) during the final evaluation on the held-out test partition. Additionally, the macro F1-score is 94.17% or 1.412% absolute difference from the second best method in published literature for this task. We also release a web-based interface built with the Streamlit framework that provides diagnostic information, including treatment and prevention guidelines, within one second of input image upload.